AERIAL IMAGE MATCHING BASED ON ZERO-CROSSINGS
Jia Zong
Jin-Cheng Li
Toni Schenk
Department of Geodetic Science and Surveying
The Ohio State University, Columbus, Ohio 43210-1247
USA
Commission III
ABSTRACT
One of the basic tasks in digital photogrammetry is to find conjugate points in a stereo pair and to reconstruct the 3-D
object space (DEM). Edges play an important role in that they may indicate breaklines in the surface. We use the LoG
operator to extract edges (zero-crossings). In this paper the problem of matching zero-crossings is addressed. Zero-crossings
computed from one image are matched with area-based method. A hierarchical matching approach is adopted by the use of
both, interpolated disparity maps at each level of the image pyramid, and knowledge from image analysis at very high level of
image pyramid. The method is particularly suited for matching aerial images for the purpose of restructing surfaces of urban
areas.
KEY WORDS: Zero-crossing, Correspondent point, Figural Continuity, Disparity Interpolation, Image Analysis.
1. INTRODUCTION
One of the major research areas in digital photogrammetry is
image matching for reconstructing the three-dimensional sur-
face of the object space. This process involves a fundamental
problem of stereo vision: to find corresponding points in an
stereo-pair. Once correspoinding points are determined their
three-dimensional positions can be easily computed, and the
surface is obtained from matched points by interpolation.
Two methods are commonly used in image matching: area-
based image matching and feature-based image matching.
Aera-based matching is predominantly used for the object
space (DEM). Here, the corresponding points are found
by comparing the gray levels of correponding areas (image
patches) in a image stereo-pair. This approach is favored
in photogrammetry because of its high accuracy potential.
However, there are several critical factors that need special
consideration in area-based matching. For example,
e good approximations for the corresponding image
patches are required
e maíching in flat area or of sharp relief changes is ex-
tremely hard and it produces bad results. Both cases
usually occur in urban aerial images
e recovering the surface, especially in urban areas, from
randomly distributed matched points is difficult
e the reliablity control of the matching is low
e computations are intensive
Some of these problems are avoided in feature-based match-
ing. Here, properties (features) derived from the gray lev-
els are matched, rather than gray levels themselves. This
method usually proceeds in two steps, the first being a lo-
cal similarity matching such as comparing the parameters of
detected features, and the second being a global matching
such as checking continuity constraints. Features detected
144
monocularly may differ and may include spurious data due
to differences in reflectance which are not caused by the sur-
face shape. This problem is quite acute in large-scale aerial
images of urban areas. Another point to bear in mind is that
matched features (e.g. edges) do not necessarily consist of
conjugate points. In general, feature-based matching is more
robust and less computationally intensive. But most impor-
tant, matched features are more meaningful than randomly
matched points if it comes to automatically analyzing image.
The motivation for this research is to combine the merits of
both area-based and feature-based matching methods. First,
edges or zero-crossings (ZC) are detected as features. The
edges are more likely to represent prominent features of the
surface, such as breaklines. Instead of matching edges as en-
tities as described in [Schenk et. al. 1991], here we match
every point of an edge by correlation. A match is accepted
if it satisfies epipolar geometry and figural continuity con-
straints. This strategy proved to be quite successful [Li et.
al. 1990]. In order to cope with urban areas where corre-
lation must be applied with caution, we have modified the
strategy by including a surface analysis step in the hierar-
chical matching scheme. At each level of the image pyramid
an interpolated disparity constraint map is generated which
provides the necessary approximations for the next level of
matching. Knowledge gained from previous levels is used
to guide matching in the subsquent level of image pyramid.
With this new strategy the success rate of matching aerial
images of complex urban scenes is greatly improved.
2. FEATURE EXTRACTION
Detecting zero-crossings as features for matching was first
proposed by Marr and Poggio [Marr and Poggio, 1979] on the
basis of a computational theory on the human stereo vision.
Mathematically, zero-crossings are obtained by applying the
convolution operator V2G over the image f(z,y) as
where
a Gau
functi
implic
vanta
and d
ized i
f'(æ,:
avera,
are t]
gray |
scribe
the g:
by an
LoG «
curve
the d
boun
Edge:
ings (
surfa
and r
eters
size c
and t
tion.
a thr
of th
are t
obtai
As a
curve
The
Like
try i
[Cho
imag
meth
are d
in th
is foi
right